PRISM-Astro/h2f_desi_ddf_cosmos_fixed

Description

Spectroscopic galaxy sample of the COSMOS Deep Drilling Field, the field the Vera C. Rubin Observatory will observe with roughly twice the cadence of the other LSST DDFs. Redshifts come from DESI Data Release 1 and images are multi-resolution cutouts of the DESI Legacy Imaging Surveys DR10 served through hips2fits by CDS.

It contains 66,373 galaxies, one row each. Every cutout is centered on the galaxy itself, so there is no augmentation and no host_pos label: the galaxy sits at the center of every image by construction. This makes it the DESI counterpart of h2f_ps1_pasquet, built with the same pipeline and the same conventions, but on a different field, a different survey and a redshift range almost five times wider.

Sample selection

Galaxies were taken from desi_dr1.zpix inside a cone of radius 1.6 deg centred on RA 150.1060, Dec +2.2260 — the median pointing of the 41 DESI tiles dedicated to COSMOS. That radius is the field of view of the Mayall with DESI's corrector: 3.2 deg across, 8.04 deg², which is what a single DESI tile covers.

An earlier version of this dataset used the 1.7481 deg circle matching LSSTCam's 9.6 deg² field of view instead. That is slightly wider than DESI's own pointing, and the outer annulus showed it: source density fell from ~9,000 to ~2,900 per deg² beyond 1.5 deg, and what remained there was mostly filled in by the wide main survey rather than the dedicated COSMOS observations. Cutting at the DESI field of view removes 5,374 galaxies and leaves a spatially complete sample. Only primary spectra with a reliable redshift were kept (zcat_primary, zwarn = 0, spectype = 'GALAXY'), and each was crossmatched to its nearest source in ls_dr10.tractor_s within 1.5 arcsec.

From the 83,620 rows that survived the crossmatch, three cuts were applied before downloading any image:

• 4,498 galaxies whose ls_id appeared more than once, since several DESI fibers can land on the same imaging source. The closest match was kept.
• 4,259 sources of Tractor type PSF or DUP, which carry shape_r = 0 and have no measurable morphology.
• 121 galaxies with a non-physical redshift, z <= 0.

A further 5,374 fall outside DESI's own field of view, as described above.

A further 2,995 were dropped after download because at least one pixel came back as 0.0 in some band, marking a gap in HiPS coverage. Those gaps are real holes in the survey rather than transient failures: re-requesting the same cutout recovered none of the 447 cases that had a whole (level, band) plane empty. No image in this dataset contains a single zero pixel.

Cross-validation

The fold column holds a 5-fold split stratified by spectroscopic redshift. Redshift is discretized into 764 uniform bins using the same bin width as h2f_ps1_pasquet (0.4 / 180 = 0.00222), extended to cover this field's much wider redshift range. The 40 galaxies falling in bins with fewer than 5 members carry fold = 0 and belong to training only, since they cannot be stratified. They sit almost entirely at the high-redshift tail, so validation metrics say nothing about the z > 1.5 regime even though the model sees those examples.

Image format

image has shape (66373, 5, 4, 30, 30), which maps to (examples, resolution levels, photometric bands, height, width). Bands are griz in that order — four bands, not the five of the PanSTARRS counterpart. Level L has a pixel scale of 0.262 * 2^L arcsec/px, so level 0 spans 7.86 arcsec and level 4 spans 125.76 arcsec around the galaxy.

Target group

group labels each galaxy as BGS, LRG, ELG or NON, following the categorisation of Zhou et al. 2024 [3]: DESI's main-survey target selection is re-applied to the photometry rather than read back from the target bitmasks. The official desitarget implementations (isBGS, isLRG, isELG) are used with south=True, since this field lies entirely in DECaLS; BGS is the union of its bright, faint and wise sub-selections, and the 628 galaxies satisfying more than one cut are assigned in the order BGS > LRG > ELG.

The distinction matters. The bitmasks record why DESI observed a galaxy, which is undefined for the special programmes that make up roughly half of this field — they leave desi_target at zero, so almost half the sample would be unclassifiable. Re-applying the cuts asks instead what the galaxy's colours make it, which is defined for every source.

group	galaxies	median z	median r
BGS	10,274	0.220	19.33
LRG	3,798	0.810	21.88
ELG	10,420	1.098	22.96
NON	41,881	0.505	21.52

survey is kept alongside it. Depth is far from uniform across this field and it varies by campaign, not by position: the SV1 tiles reach a median exposure of 13,955 s, against ~1,800–2,050 s for sv3, special and main. Cut on survey rather than on sky position if a uniform-depth subsample is needed.

Ellipse convention

ellipse_phi is not the raw catalog angle. Legacy Surveys stores ellipticity as a complex number, e = (a-b)/(a+b) * exp(2i*phi), from which the position angle is 0.5 * atan2(shape_e2, shape_e1). That angle is defined against the sky axes, and drawing it on a pixel array requires a further 90 degree rotation, verified against second-moment measurements of the cutouts themselves. This column already carries that rotation, wrapped back into [-90, 90), so it can be fed directly to a profile generator in the same way rSerPhi is used in h2f_ps1_pasquet. The raw components are kept in shape_e1 and shape_e2 for anyone who needs the catalog convention.

Galaxies of type REX are round by construction: their ellipticity is fixed at zero, so ellipse_ab = 1 and ellipse_phi = 0. That zero is a convention, not a measurement — a circle has no orientation. Use type to mask them.

Columns

Column	Type	Description
Image
image	(5,4,30,30) float32	Multi-resolution Legacy Surveys DR10 griz cutout, centered on the galaxy.
Identification
ls_id	int64	Legacy Surveys DR10 identifier of the galaxy, unique in this dataset. Packs RELEASE, BRICKID and OBJID into one integer.
targetid	int64	DESI identifier of the spectrum the redshift comes from. The key back to desi_dr1 tables and to SPARCL.
sep_arcsec	float32	Angular separation of the DESI–Legacy Surveys crossmatch, in arcsec. Median 0.06, at most 1.496.
Sky position
ra	float32	Right ascension of the galaxy, in degrees. This is the center of the cutout.
dec	float32	Declination of the galaxy, in degrees. This is the center of the cutout.
Redshift
z	float32	Spectroscopic redshift from DESI DR1, the regression target. Ranges from 0 to 1.697, with a median of 0.534.
z_error	float32	Uncertainty reported by the DESI spectroscopic pipeline.
z_class	int64	Redshift bin, 0 to 763, of width 0.00222. Used to stratify the folds and as the classification target.
group	string	Target category from the photometric selection: BGS, LRG, ELG or NON. See the section above.
survey	string	DESI campaign the spectrum comes from: sv1, sv3, special or main. Proxy for exposure depth.
Photometry
mag_g	float32	AB magnitude in the g band, as observed.
mag_r, mag_i, mag_z	float32	The same for r, i and z.
dered_mag_g	float32	AB magnitude in g corrected for Galactic extinction.
dered_mag_r, _i, _z	float32	The same for r, i and z. A handful of sources have no magnitude where the Tractor flux is negative.
Extinction
ebv	float32	Reddening E(B−V) from the SFD dust map.
mw_transmission_g, _r, _i, _z	float32	Galactic transmission per band, in linear units between 0 and 1. Multiply the pixels by the inverse to correct for extinction.
extinction_g, _r, _i, _z	float32	The same extinction expressed in magnitudes, -2.5*log10(mw_transmission).
Morphology — Tractor fit, all bands jointly
type	string	Model the Tractor selected: REX round exponential, EXP exponential, DEV de Vaucouleurs, SER free Sérsic. All are Sérsic profiles; the label says how the index was treated.
shape_r	float32	Half-light radius of the fitted model, in arcsec. Not band-specific: the Tractor fits one morphology across all bands.
shape_e1, shape_e2	float32	Raw ellipticity components as the catalog stores them. Both exactly 0 for REX.
ellipse_radius	float32	Semi-major axis of the fitted ellipse, in arcsec. Equal to shape_r; the minor axis is ellipse_radius * ellipse_ab.
ellipse_ab	float32	Axis ratio b/a, between 0 and 1. Derived as (1-|e|)/(1+|e|) with |e| = hypot(shape_e1, shape_e2).
ellipse_phi	float32	Position angle of the fitted ellipse, in degrees over [-90, 90), already rotated to the pixel frame. See the ellipse convention section above.
Cross-validation
fold	int64	Validation fold, 1 to 5, stratified by z_class. A value of 0 means the galaxy is used for training in every split.

References

[1] DESI Collaboration 2025, Data Release 1 of the Dark Energy Spectroscopic Instrument. arXiv:2503.14745

[2] Dey, A., Schlegel, D. J., Lang, D., et al. 2019, Overview of the DESI Legacy Imaging Surveys, The Astronomical Journal, 157, 168. doi:10.3847/1538-3881/ab089d

[3] Zhou, X., Li, N., Zou, H., et al. 2024, Estimating Photometric Redshifts for Galaxies from the DESI Legacy Imaging Surveys with Bayesian Neural Networks Trained by DESI EDR. arXiv:2412.02390

@article{Dey_2019,
  doi = {10.3847/1538-3881/ab089d},
  url = {https://doi.org/10.3847/1538-3881/ab089d},
  year = {2019},
  month = {apr},
  publisher = {American Astronomical Society},
  volume = {157},
  number = {5},
  pages = {168},
  author = {Dey, Arjun and Schlegel, David J. and Lang, Dustin and others},
  title = {Overview of the DESI Legacy Imaging Surveys},
  journal = {The Astronomical Journal},
}